{"id":141,"date":"2026-04-30T17:57:00","date_gmt":"2026-04-30T09:57:00","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/?p=141"},"modified":"2026-04-23T17:59:07","modified_gmt":"2026-04-23T09:59:07","slug":"%e4%bb%8e%e9%87%8d%e5%a4%8d%e5%8a%b3%e5%8a%a8%e5%88%b0%e6%99%ba%e8%83%bd%e8%ae%b0%e5%bf%86%ef%bc%9akv-cache%e4%b8%8epagedattention%e6%b7%b1%e5%ba%a6%e8%a7%a3","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/04\/30\/%e4%bb%8e%e9%87%8d%e5%a4%8d%e5%8a%b3%e5%8a%a8%e5%88%b0%e6%99%ba%e8%83%bd%e8%ae%b0%e5%bf%86%ef%bc%9akv-cache%e4%b8%8epagedattention%e6%b7%b1%e5%ba%a6%e8%a7%a3\/","title":{"rendered":"\u4ece\u201c\u91cd\u590d\u52b3\u52a8\u201d\u5230\u201c\u667a\u80fd\u8bb0\u5fc6\u201d\uff1aKV Cache\u4e0ePagedAttention\u6df1\u5ea6\u89e3\u6790"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u5199\u5728\u524d\u9762<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u524d\u9762\u7684\u516d\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u5b8c\u6574\u62c6\u89e3\u4e86Transformer\u7684\u9759\u6001\u67b6\u6784\u2014\u2014\u4eceTokenizer\u5230Embedding\uff0c\u4eceFFN\u5230Attention\uff0c\u4ece\u6b8b\u5dee\u8fde\u63a5\u5230\u5f52\u4e00\u5316\u3002\u4f46\u81f3\u6b64\uff0c\u6211\u4eec\u89e3\u7b54\u7684\u8fd8\u53ea\u662f\u201c\u6a21\u578b\u957f\u4ec0\u4e48\u6837\u201d\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u8fd9\u7bc7\u6587\u7ae0\u5f00\u59cb\uff0c\u6211\u4eec\u8981\u628a\u76ee\u5149\u4ece\u201c\u9759\u6001\u7ed3\u6784\u201d\u8f6c\u5411<strong>\u201c\u52a8\u6001\u8fd0\u884c\u201d<\/strong>\u2014\u2014\u5f53\u6a21\u578b\u771f\u6b63\u5f00\u59cb\u9010\u5b57\u751f\u6210\u6587\u672c\u65f6\uff0c\u5e95\u5c42\u5230\u5e95\u53d1\u751f\u4e86\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u4e00\u4e2a\u660e\u660e\u53ea\u67097B\u53c2\u6570\u7684\u6a21\u578b\uff0c\u8dd1\u8d77\u6765\u5374\u9700\u8981\u51e0\u5341GB\u7684\u663e\u5b58\uff1f\u4e3a\u4ec0\u4e48\u540c\u4e00\u4e2a\u53e5\u5b50\u91cc\u7684\u8bcd\uff0c\u6a21\u578b\u8981\u53cd\u590d\u8ba1\u7b97\u90a3\u4e48\u591a\u6b21\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u95ee\u9898\u7684\u7b54\u6848\uff0c\u6307\u5411\u4e00\u4e2a\u8d2f\u7a7fLLM\u63a8\u7406\u59cb\u7ec8\u7684\u5173\u952e\u673a\u5236\u2014\u2014<strong>KV Cache<\/strong>\u3002\u800c\u771f\u6b63\u8ba9KV Cache\u7684\u5185\u5b58\u7ba1\u7406\u4ece\u201c\u7c97\u653e\u201d\u8d70\u5411\u201c\u7cbe\u7ec6\u201d\u7684\u9769\u547d\u6027\u6280\u672f\uff0c\u53eb\u505a <strong>PagedAttention<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8bf4KV Cache\u662fLLM\u63a8\u7406\u7684\u201c\u53d1\u52a8\u673a\u201d\uff0c\u90a3\u4e48PagedAttention\u5c31\u662f\u4e3a\u8fd9\u53f0\u53d1\u52a8\u673a\u88c5\u4e0a\u7684\u201c\u9ad8\u6548\u71c3\u6cb9\u7cfb\u7edf\u201d\u3002\u4eca\u5929\u8fd9\u7bc7\u6587\u7ae0\uff0c\u6211\u4eec\u5c06\u6df1\u5165\u62c6\u89e3\u8fd9\u4e24\u4e2a\u7d27\u5bc6\u76f8\u5173\u7684\u6280\u672f\uff0c\u8bb2\u6e05\u695a\u5b83\u4eec\u4e3a\u4ec0\u4e48\u5982\u6b64\u91cd\u8981\uff0c\u4ee5\u53ca\u5b83\u4eec\u5982\u4f55\u5171\u540c\u63a8\u52a8\u4e86\u6574\u4e2aLLM\u670d\u52a1\u9886\u57df\u7684\u6548\u7387\u9769\u547d\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001KV Cache\uff1aLLM\u63a8\u7406\u7684\u201c\u53d1\u52a8\u673a\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1.1 \u81ea\u56de\u5f52\u751f\u6210\u7684\u201c\u6548\u7387\u9677\u9631\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8981\u7406\u89e3KV Cache\u7684\u4ef7\u503c\uff0c\u9996\u5148\u8981\u7406\u89e3\u5b83\u8981\u89e3\u51b3\u7684\u6839\u672c\u95ee\u9898\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">LLM\u751f\u6210\u56de\u7b54\u7684\u8fc7\u7a0b\uff0c\u53eb\u505a<strong>\u81ea\u56de\u5f52\u751f\u6210<\/strong>\u2014\u2014\u4e00\u4e2a\u5b57\u4e00\u4e2a\u5b57\u5730\u5f80\u5916\u8e66\u3002\u6bcf\u751f\u6210\u4e00\u4e2a\u65b0Token\uff0c\u6a21\u578b\u90fd\u8981\u628a\u5b83\u653e\u5230\u6574\u4e2a\u5e8f\u5217\u7684\u672b\u5c3e\uff0c\u91cd\u65b0\u5bf9\u6574\u4e2a\u5e8f\u5217\u505a\u4e00\u904dAttention\u8ba1\u7b97\u3002\u7528\u6570\u636e\u6d41\u6765\u76f4\u89c2\u5730\u770b\uff1a\u751f\u6210\u7b2c1\u4e2a Token\uff0c\u6a21\u578b\u5904\u74061\u4e2a\u4f4d\u7f6e\uff1b\u751f\u6210\u7b2c2\u4e2a Token\uff0c\u6a21\u578b\u5904\u74062\u4e2a\u4f4d\u7f6e\uff1b\u751f\u6210\u7b2c n \u4e2a Token\uff0c\u6a21\u578b\u5904\u7406 n \u4e2a\u4f4d\u7f6e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u610f\u5473\u7740\uff0c\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e00\u4e2a\u65e9\u671fToken\uff0c\u90fd\u4f1a\u88ab\u53cd\u590d\u5904\u7406\u5f88\u591a\u6b21\u3002\u4e00\u4e2a\u4f4d\u4e8ePrompt\u5f00\u5934\u7684\u8bcd\uff0c\u53ef\u80fd\u4f1a\u88ab\u8ba1\u7b97\u51e0\u767e\u751a\u81f3\u4e0a\u5343\u6b21\u2014\u2014\u800c\u4e14\u6bcf\u4e00\u6b21\u7684\u8ba1\u7b97\u7ed3\u679c\uff0c\u548c\u4e0a\u4e00\u6b21\u76f8\u6bd4\u6ca1\u6709\u4efb\u4f55\u53d8\u5316\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u91cd\u590d\u8ba1\u7b97\u5e26\u6765\u7684\u5f00\u9500\u662f\u5de8\u5927\u7684\uff1a\u603b\u8ba1\u7b97\u91cf\u4ee5\u5e8f\u5217\u957f\u5ea6<strong>\u5e73\u65b9\u7ea7\u522b<\/strong>\u589e\u957f\uff08O(n\u00b2)\uff09\uff0c\u5e8f\u5217\u8d8a\u957f\uff0c\u8ba1\u7b97\u4ee3\u4ef7\u8d8a\u6c89\u91cd\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.2 Attention\u4e2d\u6700\u5177\u542f\u53d1\u7684\u89c2\u5bdf<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728Transformer\u7684Self-Attention\u4e2d\uff0c\u6bcf\u4e2aToken\u88ab\u7528\u6765\u8ba1\u7b97\u4e09\u4e2a\u5411\u91cf\uff1aQuery\uff08Q\uff09\u3001Key\uff08K\uff09\u548cValue\uff08V\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u4e2a\u5173\u952e\u7684\u89c2\u5bdf\u662f\uff1a<strong>\u5bf9\u4e8e\u4e00\u4e2a\u5df2\u7ecf\u751f\u6210\u51fa\u6765\u7684\u201c\u5386\u53f2Token\u201d\u800c\u8a00\uff0c\u5b83\u7684Key\u5411\u91cf\u548cValue\u5411\u91cf\uff0c\u5728\u540e\u7eed\u7684\u751f\u6210\u8fc7\u7a0b\u4e2d\u4e0d\u4f1a\u518d\u53d1\u751f\u4efb\u4f55\u53d8\u5316\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4ec0\u4e48\u5462\uff1f\u56e0\u4e3aKey\u548cValue\u53ea\u4f9d\u8d56\u4e8e\u8fd9\u4e2aToken\u672c\u8eab\u7684Embedding\u2014\u2014\u800c\u4e00\u4e2a\u5386\u53f2Token\u7684Embedding\u662f\u56fa\u5b9a\u7684\u3002\u65e0\u8bba\u672a\u6765\u518d\u751f\u6210\u591a\u5c11\u4e2a\u65b0\u8bcd\uff0c\u5b83\u7684K\u548cV\u90fd\u4e0d\u4f1a\u53d8\u3002\u552f\u4e00\u4f1a\u968f\u7740\u5e8f\u5217\u63a8\u8fdb\u800c\u4e0d\u65ad\u66f4\u65b0\u7684\u662f\u65b0Token\u7684Query\uff0c\u4f46\u5b83\u4e0d\u9700\u8981\u548cK\u3001V\u4e00\u8d77\u201c\u65e7\u4e8b\u91cd\u63d0\u201d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.3 \u7f13\u5b58\u51b3\u5b9a\u6027\u7684\u601d\u60f3\uff1a\u7b97\u5b8c\u5c31\u5b58\u8d77\u6765\uff0c\u522b\u6254<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\uff0cKV Cache\u7684\u601d\u60f3\u662f\u98a0\u8986\u6027\u7684\uff1a<strong>\u4e0d\u8981\u628a\u90a3\u4e9b\u8f9b\u8f9b\u82e6\u82e6\u7b97\u51fa\u6765\u7684K\u548cV\u6254\u6389\u4e86\uff01\u628a\u5b83\u4eec\u7f13\u5b58\u8d77\u6765\uff01<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728Prefill\u9636\u6bb5\uff0c\u6a21\u578b\u63a5\u6536\u7528\u6237\u8f93\u5165Prompt\u7684\u6240\u6709Token\uff0c\u9884\u5148\u8ba1\u7b97\u5e76\u7f13\u5b58\u5b83\u4eec\u7684K\u548cV\u5411\u91cf\u3002\u968f\u540e\u8fdb\u5165\u89e3\u7801\u9636\u6bb5\uff0c\u6bcf\u751f\u6210\u4e00\u4e2a\u65b0Token\uff0c\u6a21\u578b\u53ea\u9700\u8981<strong>\u4e3a\u65b0Token\u8ba1\u7b97\u4e00\u4efd\u65b0\u7684K\u548cV<\/strong>\uff0c\u7136\u540e\u7528\u5b83\u7684Q\u53bbCache\u91cc\u53d6\u6240\u6709\u5386\u53f2K\u8ba1\u7b97\u5206\u6570\u3002\u6574\u4e2a\u8fc7\u7a0b\uff0c\u65e7Token\u7684Key\/Value\u5b8c\u5168\u4e0d\u9700\u8981\u4ece\u5934\u8ba1\u7b97\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6548\u679c\u662f\u663e\u8457\u7684\uff1a<strong>\u8ba1\u7b97\u590d\u6742\u5ea6\u76f4\u63a5\u4eceO(n\u00b2)\u964d\u5230\u4e86O(n)<\/strong> \u3002\u6bcf\u751f\u6210\u4e00\u4e2a\u65b0Token\uff0c\u53ea\u9700\u8981\u5e38\u6570\u65f6\u95f4\u7684\u8fd0\u7b97\uff0c\u5b9e\u9645\u63a8\u7406\u901f\u5ea6\u53ef\u80fd\u63d0\u5347<strong>3\u52305\u500d<\/strong>\uff0c\u987a\u5e8f\u8d8a\u957f\u6548\u76ca\u8d8a\u660e\u663e\u3002\u5728128K\u957f\u5e8f\u5217\u7684\u6781\u7aef\u60c5\u51b5\u4e0b\uff0c\u751a\u81f3\u53ef\u80fd\u5feb<strong>10\u523020\u500d<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.4 KV Cache\u7684\u8ba1\u7b97\u4ee3\u4ef7\u6709\u591a\u5927\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u5929\u4e0b\u6ca1\u6709\u514d\u8d39\u7684\u5348\u9910\u3002KV Cache\u867d\u7136\u5927\u5e45\u964d\u4f4e\u4e86<strong>\u8ba1\u7b97\u91cf<\/strong>\uff0c\u4f46\u5b83\u628a\u4ee3\u4ef7\u8f6c\u79fb\u5230\u4e86<strong>\u663e\u5b58\u5360\u7528<\/strong>\u4e0a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">K\u548cV\u5411\u91cf\u7684\u7f13\u5b58\u9a7b\u7559\u5728GPU\u663e\u5b58\uff08HBM\uff09\u4e2d\u3002KV Cache\u7684\u663e\u5b58\u6d88\u8017\u9075\u5faa\u4ee5\u4e0b\u516c\u5f0f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5bf9\u4e8e\u6bcf\u4e2aToken\uff0c\u6bcf\u5c42\u3001\u6bcf\u4e2a\u5934\uff0c\u90fd\u9700\u8981\u5b58\u50a8\u4e00\u4e2a K \u5411\u91cf\u548c\u4e00\u4e2a V \u5411\u91cf\u3002<\/li>\n\n\n\n<li>\u6279\u5904\u7406\u5927\u5c0f\uff08Batch Size\uff09\u518d\u4e58\u4e0a\u540c\u65f6\u5904\u7406\u7684\u8bf7\u6c42\u6570\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6839\u636e\u5b9e\u9645\u6d4b\u91cf\uff0c\u5f53\u4e0a\u4e0b\u6587\u8d85\u8fc78000 Token\u65f6\uff0cKV Cache\u7684\u663e\u5b58\u5360\u7528\u5c31\u4f1a\u8d85\u8fc7\u6a21\u578b\u6743\u91cd\u3002\u52301M\u4e0a\u4e0b\u6587\u65f6\uff0cKV Cache\u53ef\u80fd\u4f1a\u5403\u6389<strong>\u8d85\u8fc795%\u7684\u603b\u663e\u5b58<\/strong>\u3002\u66f4\u76f4\u89c2\u7684\u4f53\u611f\u662f\uff0c\u5904\u7406<strong>\u767e\u4e07\u7ea7\u4e0a\u4e0b\u6587\u7684\u63a8\u7406\uff0c\u5176KV Cache\u53ef\u80fd\u5360\u7528\u6570\u767eGB\u7684\u663e\u5b58<\/strong>\u3002\u56e0\u6b64\uff0cKV Cache\u7684\u8fd9\u79cd\u65e0\u5e8f\u6269\u5f20\uff0c\u6b63\u5728\u6210\u4e3a\u5927\u578b\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u90e8\u7f72\u7684<strong>\u6838\u5fc3\u74f6\u9888<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.5 \u4f20\u7edf\u5185\u5b58\u5206\u914d\u7684\u4e09\u5b97\u7f6a<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728PagedAttention\u51fa\u73b0\u4e4b\u524d\uff0c\u63a8\u7406\u6846\u67b6\uff08\u6bd4\u5982\u6700\u539f\u59cb\u7684Hugging Face Transformers\u5b9e\u73b0\uff09\u5bf9\u5f85KV Cache\u7684\u65b9\u5f0f\u975e\u5e38\u7c97\u653e\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u9884\u7559\u5f0f\u5206\u914d<\/strong>\uff1a\u7cfb\u7edf\u4f1a\u4e3a\u6bcf\u4e2a\u63a8\u7406\u8bf7\u6c42\uff0c\u5728\u663e\u5b58\u91cc\u9884\u5148\u5212\u5b9a\u4e00\u5757\u5de8\u5927\u7684\u3001\u8fde\u7eed\u7684\u7269\u7406\u7a7a\u95f4\uff0c\u5176\u5927\u5c0f\u76f4\u63a5\u7b49\u4e8e\u201c\u6700\u5927\u53ef\u80fd\u751f\u6210\u7684\u5e8f\u5217\u957f\u5ea6\u201d\u3002\u7edd\u5927\u90e8\u5206\u9884\u7559\u597d\u7684\u663e\u5b58\u5176\u5b9e\u4ec0\u4e48\u90fd\u6ca1\u6709\u5199\uff0c\u4f46\u7cfb\u7edf\u4e0d\u6562\u590d\u7528\uff0c\u9020\u6210\u4e86\u6b8b\u9177\u7684<strong>\u5185\u90e8\u788e\u7247<\/strong>\u3002<\/li>\n\n\n\n<li><strong>\u7269\u7406\u8fde\u7eed\u6027<\/strong>\uff1a\u4f20\u7edfAttention Kernel\u8981\u6c42KV Cache\u5fc5\u987b\u5728\u663e\u5b58\u4e2d\u8fde\u6210\u4e00\u7247\u3002\u5728\u5904\u7406\u5927\u91cf\u5e76\u53d1\u7684\u3001\u957f\u5ea6\u5404\u5f02\u7684\u8bf7\u6c42\u65f6\uff08\u6709\u7684\u53ea\u804a\u4e24\u53e5\uff0c\u6709\u7684\u8981\u957f\u7bc7\u5927\u8bba\uff09\uff0c\u8fd9\u79cd\u5206\u914d\u65b9\u5f0f\u50ac\u751f\u4e86\u66f4\u52a0\u6076\u52a3\u7684<strong>\u5916\u90e8\u788e\u7247<\/strong>\u3002\u5c0f\u5757\u65e0\u6548\u5185\u5b58\u5939\u5728\u4e0d\u540c\u957f\u5ea6\u7684\u6d3b\u8dc3Cache\u4e4b\u95f4\uff0c\u6210\u4e86\u65e0\u4eba\u80fd\u7528\u7684\u201c\u5b64\u5c9b\u201d\u3002<\/li>\n\n\n\n<li><strong>\u9759\u6001\u72ec\u5360<\/strong>\uff1a\u6bcf\u4e2a\u8bf7\u6c42\u72ec\u4eab\u4e00\u6574\u5757\u9884\u7559\u533a\u57df\uff0c\u76f4\u5230\u751f\u6210\u7ed3\u675f\u624d\u6574\u4f53\u91ca\u653e\uff0c\u65e0\u6cd5\u52a8\u6001\u4f38\u7f29\u6216\u5171\u4eab\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u5b9e\u8df5\u4e2d\uff0c\u8fd9\u79cd\u201c\u7c97\u653e\u5f0f\u7ecf\u8425\u201d\u7684\u6076\u679c\u662f\u663e\u5b58\u4e2d\u5145\u65a5\u7740\u5927\u91cf\u7684\u65e0\u6548\u788e\u7247\u3002\u6709\u5206\u6790\u6307\u51fa\uff0c\u4f20\u7edf\u5185\u5b58\u7ba1\u7406\u65b9\u5f0f\u4e0b\uff0c\u9ad8\u8fbe80%\u7684\u9884\u7559\u663e\u5b58\u5728\u5927\u90e8\u5206\u65f6\u5019\u90fd\u662f\u7a7a\u7f6e\u7684\u3002\u6a21\u578b\u663e\u5b58\u867d\u5927\uff0c\u53ef\u5e76\u53d1\u5904\u7406\u591a\u4e2a\u8bf7\u6c42\u7684\u80fd\u529b\uff08\u541e\u5410\u91cf\uff09\u5374\u88ab\u4e25\u91cd\u675f\u7f1a\u4e86\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001PagedAttention\uff1a\u89e3\u5f00KV Cache\u7684\u201c\u5185\u5b58\u6b7b\u7ed3\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 \u9769\u547d\u6027\u7684\u7075\u611f\uff1a\u7ecf\u5178\u7684\u865a\u62df\u5185\u5b58\u5206\u9875<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u9762\u5bf9\u6ee1\u76ee\u75ae\u75cd\u7684\u663e\u5b58\u788e\u7247\uff0c\u6765\u81eaUC Berkeley\u7684vLLM\u56e2\u961f\u6ca1\u6709\u63a5\u7740\u5728Attention\u516c\u5f0f\u672c\u8eab\u4e0a\u201c\u6253\u8865\u4e01\u201d\uff0c\u800c\u662f\u505a\u51fa\u4e00\u4e2a\u5927\u80c6\u7684\u8de8\u754c\u7c7b\u6bd4\uff1a<strong>\u663e\u5b58\u788e\u7247\u5316\uff1f\u8fd9\u4e0d\u662f\u64cd\u4f5c\u7cfb\u7edf\u7684\u865a\u62df\u5185\u5b58\u7ba1\u7406\u65e9\u5c31\u89e3\u51b3\u4e86\u7684\u95ee\u9898\u5417\uff1f<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u64cd\u4f5c\u7cfb\u7edf\u65e9\u5c31\u629b\u5f03\u4e86\u4e3a\u6bcf\u4e2a\u8fdb\u7a0b\u9884\u7559\u5927\u5757\u8fde\u7eed\u7269\u7406\u5185\u5b58\u7684\u505a\u6cd5\u3002\u53d6\u800c\u4ee3\u4e4b\u7684\u662f<strong>\u5206\u9875\u673a\u5236<\/strong>\uff1a\u628a\u6574\u4e2a\u5185\u5b58\u7a7a\u95f4\u5207\u5206\u6210\u4e00\u4e2a\u4e2a\u56fa\u5b9a\u5927\u5c0f\u7684\u5c0f\u5757\uff08Page\uff09\uff0c\u518d\u7528\u4e00\u5f20\u5168\u5c40\u7684\u201c\u9875\u8868\u201d\u628a\u201c\u8fdb\u7a0b\u89c6\u89d2\u4e0b\u8fde\u7eed\u7684\u903b\u8f91\u5730\u5740\u201d\uff0c\u7075\u6d3b\u5730\u6620\u5c04\u5230\u7269\u7406\u5185\u5b58\u4e2d\u4efb\u610f\u4e0d\u8fde\u7eed\u7684\u7a7a\u95f2\u9875\u4e0a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ed6\u4eec\u57282023\u5e74\u7684SOSP\u4f1a\u8bae\u4e0a\uff0c\u5c06\u8fd9\u4e2a\u5df2\u7ecf\u8fc7\u534a\u4e2a\u4e16\u7eaa\u68c0\u9a8c\u7684\u5de5\u7a0b\u54f2\u5b66\uff0c\u5b8c\u7f8e\u8fc1\u79fb\u5230\u4e86KV Cache\u7684\u7ba1\u7406\u4e0a\uff0c\u547d\u540d\u5176\u4e3a\u2014\u2014<strong>PagedAttention<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 \u533a\u5757\u5316\u7684KV Cache\uff1a\u5b83\u662f\u5982\u4f55\u8fd0\u4f5c\u7684\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">PagedAttention\u6539\u53d8\u4e86KV Cache\u7684\u57fa\u672c\u5b58\u50a8\u5355\u4f4d\u3002\u5b83\u4e0d\u518d\u6309\u201c\u6574\u4e2a\u5e8f\u5217\u201d\u5206\u914d\u5185\u5b58\uff0c\u800c\u662f\u5f3a\u5236\u5c06\u6bcf\u4e2a\u5e8f\u5217\u7684KV Cache\uff0c\u5728\u5e95\u5c42\u5207\u5206\u6210\u4e00\u4e2a\u4e2a\u56fa\u5b9a\u5927\u5c0f\u7684 <strong>Block<\/strong>\u3002\u5e38\u89c1\u7684\u5757\u5927\u5c0f\u662f <strong>16<\/strong> \u6216 <strong>32<\/strong> \u4e2aToken\u4f5c\u4e3a\u4e00\u4e2aBlock\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a\u5e8f\u5217\u5185\u90e8\u7ef4\u62a4\u4e00\u4e2a <strong>Block Table<\/strong>\uff08\u5757\u8868\uff09\u3002\u5b83\u5c06Token\u7684\u201c\u903b\u8f91\u4f4d\u7f6e\u201d\uff08\u4f8b\u5982\uff0cToken 0-15\u5728\u903b\u8f91\u5757#0\uff09\u6620\u5c04\u5230\u8fd9\u4e2a\u903b\u8f91\u5757\u5728\u5f53\u524d\u65f6\u523b\u6240\u5728\u7684\u771f\u5b9e\u201c\u7269\u7406\u5757\u201d\u7f16\u53f7\u3002\u591a\u4e2a\u8bf7\u6c42\u5185\u90e8\u4f7f\u7528\u5404\u81ea\u4e0d\u540c\u7684\u5757\u8868\u505a\u6620\u5c04\uff0c\u5b83\u4eec\u7684KV cache\u5b9e\u9645\u4e0a\u53ef\u80fd\u6563\u843d\u5728\u663e\u5b58\u7684\u5404\u4e2a\u89d2\u843d\u3002\u4f46\u4ece\u5355\u4e2a\u8bf7\u6c42\u89c6\u89d2\u770b\uff0c\u5b83\u4eec\u4f9d\u7136\u5f62\u6210\u4e86\u4e00\u4e2a\u903b\u8f91\u4e0a\u8fde\u7eed\u7684\u5e8f\u5217\u3002\u56e0\u6b64\uff0cBlock Table\u5c31\u76f8\u5f53\u4e8e\u64cd\u4f5c\u7cfb\u7edf\u7684<strong>\u9875\u8868<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 \u5185\u5b58\u6548\u7387\u7684\u9769\u547d\u6027\u63d0\u5347<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u9897\u7c92\u5ea6\u7a0d\u7c97\u4f46\u6781\u5176\u7075\u6d3b\u7684\u7ba1\u7406\uff0c\u5e26\u6765\u4e86\u81f3\u5c11\u4e09\u4e2a\u7ef4\u5ea6\u7684\u663e\u8457\u6536\u76ca\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9996\u5148\uff0c\u5185\u90e8\u788e\u7247\u51e0\u4e4e\u5f7b\u5e95\u6d88\u5931\uff1a\u7cfb\u7edf\u53ea\u4e3a\u5df2\u7ecf\u751f\u6210\u51fa\u6765\u7684Token\u5206\u914d\u4e0b\u4e00\u5757Block\uff0c\u800c\u4e14\u6bcf\u5757\u90fd\u4e0d\u5927\u3002\u4e00\u4e2a\u8bf7\u6c42\u5f53\u524d\u5269\u51e0\u4e2aToken\u6ca1\u88c5\u6ee1\uff0c\u5c31\u6d6a\u8d39\u51e0\u4e2a\u4f4d\u7f6e\uff0c\u9020\u6210\u7684\u5185\u5b58\u6d6a\u8d39\u5e73\u5747\u8fdc\u4f4e\u4e8e\u4f20\u7edf\u9884\u7559\u65b9\u6848\u768420-40%\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u6b21\uff0c\u5916\u90e8\u788e\u7247\u5f7b\u5e95\u6d88\u9664\uff1a\u56e0\u4e3a\u662f\u52a8\u6001\u6620\u5c04\uff0c\u4e0d\u8981\u6c42\u7269\u7406\u5757\u8fde\u7eed\uff0c\u6240\u4ee5\u201c\u5916\u90e8\u788e\u7247\u201d\u8fd9\u4e2a\u6982\u5ff5\u5728PagedAttention\u4f53\u7cfb\u4e0b\u4ece\u6839\u672c\u4e0a\u6d88\u5931\u4e86\u3002\u7cfb\u7edf\u53ef\u4ee5\u9ad8\u6548\u652f\u914d\u6bcf\u4e00\u5bf8\u7a7a\u95f2\u7684GPU\u5185\u5b58\uff0c\u5229\u7528\u7387\u7a7a\u524d\u63d0\u9ad8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u540e\uff0c\u6309\u9700\u52a8\u6001\u4f38\u7f29\uff1a\u968f\u7740\u5e8f\u5217\u4e0d\u65ad\u751f\u6210\u65b0\u5185\u5bb9\uff0cBlock Table\u52a8\u6001\u8ffd\u52a0\u6620\u5c04\u65b0\u7684\u7269\u7406\u5757\u3002\u5f53\u8bf7\u6c42\u5b8c\u6210\u540e\uff0c\u5757\u88ab\u6574\u4f53\u56de\u6536\u5230\u4e00\u4e2a\u5168\u5c40\u7684\u7a7a\u95f2\u5757\u6c60\u5b50\u91cc\u4f9b\u65b0\u8bf7\u6c42\u4f7f\u7528\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.4 \u7ec8\u6781\u79d8\u8bc0\uff1a\u96f6\u5197\u4f59\u7684Copy-on-Write\u5171\u4eab<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u771f\u5b9eLLM\u670d\u52a1\u573a\u666f\u4e2d\uff08\u5982RAG\u6216\u8005\u5e76\u884c\u91c7\u6837\uff09\uff0c\u4e00\u4e2a\u56e2\u961f\u5f80\u5f80\u4f1a\u9488\u5bf9\u540c\u4e00\u6bb5\u5f88\u957f\u7684\u7cfb\u7edf\u63d0\u793a\u8bcd\uff08System Prompt\uff09\uff0c\u53d1\u9001\u591a\u4e2a\u4e0d\u540c\u7684\u7528\u6237\u95ee\u9898\u3002\u5982\u679c\u4e0d\u91c7\u7528\u5171\u4eab\u673a\u5236\uff0c\u7cfb\u7edf\u5fc5\u987b\u4e3a\u6bcf\u4e2a\u5e76\u53d1\u7684\u8bf7\u6c42\u56e2\uff0c\u5355\u72ec\u5206\u914d\u5e76\u8ba1\u7b97\u540c\u4e00\u4efd\u7cfb\u7edf\u63d0\u793a\u8bcd\u7684KV Cache\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u800cPagedAttention\u901a\u8fc7\u5728Block Table\u5c42\u9762\u5b9e\u73b0<strong>\u201c\u5199\u65f6\u590d\u5236\u201d<\/strong> \u8bed\u4e49\uff0c\u4f18\u96c5\u5730\u89e3\u51b3\u4e86\u6b64\u95ee\u9898\u3002\u5f53\u4ece\u540c\u4e00\u4e2aSystem Prompt\u6d3e\u751f\u591a\u4e2a\u65b0\u8bf7\u6c42\u65f6\uff0c\u7cfb\u7edf\u53ea\u9700\u8981\u4e3a\u8fd9\u4e9b\u8bf7\u6c42\u521b\u5efa\u65b0\u7684\u903b\u8f91\u5757\u8868\uff0c\u800c\u8868\u91cc\u5927\u90e8\u5206\u6761\u76ee\u6620\u5c04\u5230\u4e86\u4e00\u4efd\u6765\u81ea\u6e90\u8bf7\u6c42\u7684<strong>\u76f8\u540c\u7269\u7406\u5757<\/strong>\u4e0a\u3002\u53ea\u6709\u5f53\u65b0\u8bf7\u6c42\u81ea\u8eab\u540e\u7eed\u751f\u6210\u7684\u72ec\u7279\u5185\u5bb9\u8d85\u51fa\u4e86\u8fd9\u4e9b\u5171\u4eab\u5757\u65f6\uff0c\u624d\u4f1a\u4e3a\u5b83\u89e6\u53d1\u65b0\u7684\u7269\u7406\u5757\u5206\u914d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7531\u4e8e\u65e0\u9700\u989d\u5916\u8ba1\u7b97\u6216\u91cd\u590d\u5b58\u50a8\uff0c\u5171\u4eab\u7ec4\u6240\u9700\u7684\u663e\u5b58\uff0c\u4e0e\u53ea\u6709\u4e00\u4e2a\u8bf7\u6c42\u7684\u60c5\u51b5\u51e0\u4e4e\u76f8\u540c\u3002vLLM\u7684\u8bc4\u4f30\u663e\u793a\uff0c\u8fd9\u9879\u7279\u6027\u6700\u9ad8\u53ef\u5c06\u5185\u5b58\u6d6a\u8d39\u8fdb\u4e00\u6b65\u964d\u4f4e\u7ea655%\uff0c\u5e76\u5e26\u6765\u81f3\u591a<strong>2.2\u500d\u7684\u541e\u5410\u91cf\u63d0\u5347<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.5 \u6fc0\u8fdb\u52a0\u901f\u80cc\u540e\u7684\u5de5\u7a0b\u590d\u6742\u6027<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u7136\uff0c\u8fd9\u79cd\u7075\u6d3b\u6027\u7684\u4ee3\u4ef7\u662f\u5e95\u5c42\u7684\u5de5\u7a0b\u590d\u6742\u5ea6\u6210\u500d\u4e0a\u5347\u3002Block Table\u5e26\u6765\u4e86\u989d\u5916\u7684\u786c\u4ef6\u8df3\u8f6c\u6307\u4ee4\uff0c\u975e\u8fde\u7eed\u5185\u5b58\u8bbf\u5b58\u4e5f\u5bf9GPU\u5e76\u884c\u6027\u6784\u6210\u8003\u9a8c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u6b64\uff0cvLLM\u5fc5\u987b\u914d\u5957\u5f00\u53d1\u4e00\u4e2a\u5168\u65b0\u7684\u3001\u9ad8\u5ea6\u4f18\u5316\u7684 <strong>CUDA Attention Kernel<\/strong> \uff0c\u4ee5\u9ad8\u6548\u5904\u7406\u8fd9\u79cd\u201c\u6563\u843d\u5728\u5404\u5904\u201d\u3001\u9760Block Table\u7d22\u5f15\u7684KV Cache\u3002\u8be5\u6838\u51fd\u6570\u4f7f\u7528\u4e86\u7cbe\u5de7\u7684\u7b56\u7565\uff08\u6bd4\u5982<strong>\u7ebf\u7a0b\u7c97\u5316\u3001\u5411\u91cf\u5316\u52a0\u8f7d\u3001Warp\u7ea7\u5e76\u884c\u5f52\u7ea6<\/strong>\uff09\uff0c\u5927\u5e45\u538b\u4f4e\u4e86\u5730\u5740\u6620\u5c04\u548c\u8c03\u5ea6\u9020\u6210\u7684\u5f00\u9500\u3002\u6b64\u5916\uff0cvLLM\u8fd8\u8bbe\u8ba1\u4e86 <strong>v2\u5206\u533a\u5f0f\u6838\u51fd\u6570<\/strong>\uff0c\u4e13\u95e8\u7528\u4e8e\u5e94\u4ed8\u4e0a\u4e0b\u6587\u6781\u957f\uff08\u5982128K Token\uff09\u7684\u573a\u666f\uff0c\u907f\u514d\u5355\u6838\u51fd\u6570\u5c40\u90e8\u663e\u5b58\u6ea2\u51fa\u5bfc\u81f4\u7684\u5d29\u6e83\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.6 \u4ece\u8bba\u6587\u5230\u843d\u5730\uff1aPagedAttention\u5f00\u6e90\u751f\u6001\u4e0e\u73b0\u5b9e\u5f71\u54cd<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">PagedAttention\u4ece\u6839\u672c\u4e0a\u6539\u53d8\u4e86\u5927\u6a21\u578b\u843d\u5730\u670d\u52a1\u7684\u6210\u672c\u66f2\u7ebf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u642d\u8f7dPagedAttention\u7684vLLM\u63a8\u7406\u6846\u67b6\u5c55\u73b0\u51fa\u4e86\u60ca\u4eba\u7684\u80fd\u6548\u4f18\u52bf\u3002\u76f8\u6bd4\u672a\u4f18\u5316\u7684Hugging Face\uff08HF\uff09\u4ee5\u53ca\u4e13\u95e8\u8bbe\u8ba1\u7684Text Generation Inference\uff08TGI\uff09\uff1a<strong>vLLM\u7684\u541e\u5410\u91cf\u6700\u9ad8\u53ef\u8fbeHF\u768424\u500d\uff0cTGI\u76843.5\u500d<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u98a0\u8986\u6027\u7684\u6548\u7387\u63d0\u5347\uff0c\u4f7f\u5f97vLLM\u548c\u80cc\u540e\u7684PagedAttention\u8fc5\u901f\u6210\u4e3a\u4e86\u6574\u4e2aLLM\u63a8\u7406\u9886\u57df\u7684\u5b9e\u9645\u6807\u51c6\u3002\u8fd1\u4e24\u5e74\uff0c\u5b83\u5df2\u4ece\u4e00\u9879\u5355\u7eaf\u7684\u5b66\u672f\u539f\u578b\uff0c\u6210\u957f\u4e3a\u4e1a\u754c\u5305\u62ecAnyscale\u3001Modal\u3001Baseten\u3001AWS\u3001Google Cloud\u7b49\u4f17\u591aLLM\u63a8\u7406\u670d\u52a1\u7684\u9ed8\u8ba4\u57fa\u77f3\u6216\u96c6\u6210\u65b9\u6848\u3002\u6beb\u4e0d\u5938\u5f20\u5730\u8bf4\uff0cPagedAttention\u7684CMO\uff08\u9996\u5e2d\u5185\u5b58\u5b98\uff09\u7ea7\u522b\u8bbe\u8ba1\uff0c\u4e3a2024-2026\u5e74\u95f4\u5927\u6a21\u578b\u5e94\u7528\u7684\u5e73\u6c11\u5316\u63d0\u4f9b\u4e86\u6700\u4e3b\u8981\u7684\u5de5\u7a0b\u63a8\u529b\u4e4b\u4e00\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u524d\u7aef\u63a2\u7d22\uff1aKV Cache\u6b63\u5728\u53d1\u751f\u4ec0\u4e48\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u804a\u5b8c\u4e86\u4f5c\u4e3a\u6838\u5fc3\u5f15\u64ce\u7684KV Cache\u548c\u5960\u5b9a\u5185\u5b58\u7ba1\u7406\u57fa\u77f3\u7684PagedAttention\uff0c\u6211\u4eec\u73b0\u5728\u5c06\u89c6\u91ce\u6295\u5411\u524d\u6cbf\u3002\u9762\u5bf9\u767e\u4e07Token\u4e0a\u4e0b\u6587\u65f6\u4ee3\u7684\u5230\u6765\uff0c\u5168\u884c\u4e1a\u90fd\u5728\u6cbf\u7740\u4e0d\u540c\u7684\u6280\u672f\u8def\u5f84\uff0c\u8bd5\u56fe\u9a6f\u670dKV Cache\u8fd9\u5934\u4ecd\u5728\u75af\u72c2\u6210\u957f\u7684\u201c\u663e\u5b58\u5de8\u517d\u201d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 \u6781\u9650\u91cf\u5316\uff1a\u8c37\u6b4cTurboQuant\u7684\u7834\u58c1\u65f6\u523b\uff082026\uff09<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u91cf\u5316\u662f\u6700\u53e4\u8001\u4e5f\u4e00\u76f4\u5728\u7528\u7684\u7597\u6cd5\u3002\u4f46\u8fc7\u53bb\uff0c\u91cf\u5316\u7cbe\u5ea6\u964d\u5230\u4e00\u5b9a\u7a0b\u5ea6\uff08\u6bd4\u59824-bit\u4ee5\u4e0b\uff09\u603b\u8981\u9762\u4e34\u4e00\u4e2a\u547d\u95e8\u2014\u2014<strong>\u7cbe\u5ea6\u635f\u5931\u5bfc\u81f4\u6a21\u578b\u8f93\u51fa\u53d8\u5dee<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">2026\u5e743\u6708\uff0c\u8c37\u6b4c\u5728ICLR\u63a8\u51fa\u7684 <strong>TurboQuant<\/strong> \u5374\u9707\u52a8\u4e86\u6574\u4e2a\u884c\u4e1a\u3002\u5b83\u58f0\u79f0\u53ef\u4ee5\u628a\u9ad8\u7cbe\u5ea6\uff08FP16\uff09\u7684KV Cache\uff0c\u60ca\u4eba\u5730<strong>\u65e0\u635f\u538b\u7f29\u81f33.5 bits<\/strong>\u2014\u2014\u538b\u7f29\u7387\u76f4\u903c<strong>6\u500d<\/strong>\uff0c\u5e76\u5b9e\u73b0<strong>8\u500d<\/strong>\u7684\u63a8\u7406\u52a0\u901f\uff0c\u5f15\u53d1\u4e86\u7f8e\u5149\u3001\u897f\u90e8\u6570\u636e\u7b49\u5185\u5b58\u5de8\u5934\u7684\u80a1\u4ef7\u9707\u52a8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">TurboQuant\u5230\u5e95\u505a\u4e86\u4ec0\u4e48\uff1f\u5b83\u7ec4\u5408\u4e86\u4e00\u5957\u6781\u5176\u806a\u660e\u7684\u53d8\u6362\u91cf\u5316\u6cd5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>PolarQuant\uff08\u6781\u5750\u6807\u91cf\u5316\uff09<\/strong>\uff1a\u653e\u5f03\u4f20\u7edf\u7684XYZ\u5750\u6807\uff0c\u5c06\u9ad8\u7ef4\u5411\u91cf\u5148\u6295\u5f71\u6210\u201c\u89d2\u5ea6\u201d\uff0c\u7531\u4e8e\u89d2\u5ea6\u5206\u5e03\u5f88\u96c6\u4e2d\uff0c\u63cf\u8ff0\u5b83\u53ea\u9700\u6781\u4f4e\u7684\u7801\u7387\u3002<\/li>\n\n\n\n<li><strong>QJL\uff08\u91cf\u5316JL\u53d8\u6362\uff09<\/strong>\uff1a\u518d\u628a\u6295\u5f71\u4ea7\u751f\u7684\u5fae\u5c0f\u504f\u5dee\u7528+1\/-1\u7684\u7b26\u53f7\u7f16\u7801\u4f5c\u4e3a\u6b8b\u5dee\u6821\u6b63\uff0c\u51e0\u4e4e\u96f6\u989d\u5916\u5f00\u9500\u5730\u6d88\u9664\u4e86\u8bef\u5dee\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u968f\u7740\u6280\u672f\u6210\u719f\uff0c\u5b83\u53ef\u80fd\u4ece\u6839\u672c\u4e0a\u6539\u53d8GPU\u786c\u4ef6\u9700\u6c42\uff0c\u8ba9\u5355\u5361\u5e26\u52a8\u6570\u767e\u4e07\u751a\u81f3\u65e0\u9650\u7684\u4e0a\u4e0b\u6587\u4e0d\u518d\u9065\u4e0d\u53ef\u53ca\u3002\u4e1a\u5185\u9884\u6d4b\uff0c\u82e5\u8be5\u6280\u672f\u672a\u6765\u5f7b\u5e95\u6210\u719f\uff0c\u5f53\u524d\u7528\u4e8eAI\u670d\u52a1\u7684\u6570\u636e\u4e2d\u5fc3GPU\u9700\u6c42\u6216\u51cf\u5c1150%\u4ee5\u4e0a\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 \u8d85\u957f\u4e0a\u4e0b\u6587\u7684\u7f13\u5b58\u9a71\u9010\u96be\u9898\uff1a2026\u7684\u591a\u5143\u89e3\u6cd5<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u4e0a\u4e0b\u6587\u7a97\u53e3\u7a81\u7834\u767e\u4e07Token\uff0c\u5355\u7eaf\u9760\u5b58\u50a8\u5df2\u65e0\u6cd5\u6ee1\u8db3\u9700\u8981\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>LookaheadKV<\/strong>\uff1a\u4e00\u79cd\u9884\u6d4b-\u9a8c\u8bc1\u5f0f\u9a71\u9010\u6cd5\u3002\u5b83\u901a\u8fc7\u5f02\u6b65\u201c\u5077\u7aa5\u201d\u672a\u6765\u53ef\u80fd\u7684\u72b6\u6001\uff0c\u6765\u51b3\u5b9a\u73b0\u5728\u54ea\u4e9bKV\u662f\u201c\u4e07\u4e07\u4e0d\u80fd\u6254\u7684\u201d\uff08\u957f\u5c3e\u4f9d\u8d56\uff09\uff0c\u5728LongBench\u4e0a\u5b9e\u73b0\u4e86\u4ec5\u9700<strong>6%\u7684\u7f13\u5b58<\/strong>\uff0c\u4f9d\u7136\u903c\u8fd199%\u5b8c\u6574\u7f13\u5b58\u7684\u6027\u80fd\u3002<\/li>\n\n\n\n<li><strong>TTKV\uff08Temporal-Tiered KV Cache\uff09<\/strong>\uff1a\u63d0\u51fa\u4e86\u201c\u65f6\u95f4\u5206\u5c42\u201d\u67b6\u6784\u3002\u521a\u751f\u6210\u7684KV\u901f\u5ea6\u5feb\u4e14\u9891\u7e41\u8bbf\u95ee\uff0c\u653e\u5feb\u901f\u5b58\u50a8\uff1b\u5386\u53f2\u53e4\u65e9\u7684KV\u901a\u8fc7\u4e00\u79cd\u5c42\u95f4\u590d\u7528\u7d22\u5f15\u8bbf\u95ee\uff0c\u653e\u6162\u901f\u5b58\u50a8\u3002\u5728128K\u4efb\u52a1\u4e0a\u8de8\u7ea7\u6d41\u91cf\u51cf\u5c11\u4e86<strong>5.94\u500d<\/strong>\u3002<\/li>\n\n\n\n<li><strong>KeyDiff<\/strong>\uff1a\u4e00\u79cd\u5b8c\u5168\u4e0d\u4f9d\u8d56\u8bad\u7ec3\u3001\u8ba1\u7b97\u4e5f\u6781\u8f7b\u7684\u9a71\u9010\u65b9\u6cd5\u3002\u5b83\u5728\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u4e2d\uff0c\u53ef\u57fa\u4e8eKey\u5411\u91cf\u4e4b\u95f4\u7684\u7eaf\u51e0\u4f55\u76f8\u4f3c\u5ea6\uff0c\u5feb\u901f\u8bc6\u522b\u5e76\u6254\u6389\u201c\u8bed\u4e49\u4e0a\u6700\u65e0\u5173\u7d27\u8981\u7684KV\u5bf9\u201d\uff0c\u7279\u522b\u9002\u5408\u8d44\u6e90\u53d7\u9650\u5c0f\u578b\u8bbe\u5907\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3.3 \u8de8\u8bf7\u6c42\u590d\u7528\uff1a\u4ece\u96f6\u8ba1\u7b97\u5230&#8221;\u5b58\u4e86\u4e00\u6b21\uff0c\u53d7\u76ca\u7ec8\u8eab&#8221;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f20\u7edf\u7684KV\u7f13\u5b58\u662f\u201c\u4e00\u6b21\u6027\u7684\u201d\uff1a\u540c\u4e00\u4e2a\u7cfb\u7edf\u63d0\u793a\u8bcd\uff08System Prompt\uff09\uff0c\u867d\u88ab\u6210\u5343\u4e0a\u4e07\u6b21\u8c03\u7528\uff0c\u6bcf\u6b21\u5374\u4ecd\u7136\u8981\u91cd\u590d\u4ece\u96f6\u6784\u5efa\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>LMCache\uff082025\uff09<\/strong> \u5f7b\u5e95\u6539\u53d8\u4e86\u8fd9\u4e2a\u8303\u5f0f\u3002\u901a\u8fc7\u5728GPU\u3001CPU\u5185\u5b58\u751a\u81f3\u672c\u5730NVMe\u78c1\u76d8\u4e0a\u5efa\u7acb<strong>\u591a\u5c42\u7ea7\u3001\u6301\u4e45\u5316\u7684KV Cache\u5b58\u50a8<\/strong>\uff0cLMCache\u4f7f\u5f97\u76f8\u540c\u6216\u76f8\u4f3c\u6587\u672c\u7684\u7f13\u5b58\uff0c\u80fd\u591f\u5728\u5b8c\u5168\u65e0\u5173\u7684\u8bf7\u6c42\u4e4b\u95f4\u667a\u80fd\u590d\u7528\u3002\u5b9e\u6d4b\u6570\u636e\u663e\u793a\uff0c\u5728\u804a\u5929\u673a\u5668\u4eba\u3001RAG\u53ca\u591a\u8f6e\u5bf9\u8bdd\u573a\u666f\u4e0b\uff0c\u4f7f\u7528LMCache\u80fd\u591f\u5c06\u9996Token\u5ef6\u8fdf\uff08TTFT\uff09\u5927\u5e45\u7f29\u51cf\uff0c\u83b7\u5f97<strong>3\u81f310\u500d\u7684\u901f\u5ea6\u63d0\u5347<\/strong>\u800c\u4e0d\u635f\u5931\u4efb\u4f55\u541e\u5410\u91cf\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.4 Spotlight Attention\uff1a\u7ed9\u957f\u4e0a\u4e0b\u6587\u753b\u4e00\u6761\u201c\u6377\u5f84\u201d\uff082025\uff09<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u4e0a\u4e0b\u6587\u957f\u8fbe512K\u751a\u81f3\u767e\u4e07Token\uff0c\u201c\u67e5\u627e\u76f8\u4f3cKey\u201d\u8fd9\u4e2a\u52a8\u4f5c\uff0c\u4e5f\u53ef\u80fd\u53d8\u6210\u62d6\u57ae\u63a8\u7406\u7684\u4e3b\u72af\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Spotlight Attention\uff082025\uff09<\/strong> \u4e3a\u6b64\u63d0\u4f9b\u4e86\u4e00\u4e2a\u601d\u8def\uff1a\u5229\u7528\u65b0\u578b\u7684<strong>\u975e\u7ebf\u6027\u54c8\u5e0c\u51fd\u6570<\/strong>\uff0c\u4e3a\u6bcf\u4e00\u4e2aKey\u548cQuery\u9884\u5148\u8ba1\u7b97\u51fa\u4e00\u4e2a<strong>\u8d85\u9ad8\u6548\u7387\u7684\u68c0\u7d22\u7d22\u5f15<\/strong>\u3002\u5b83\u901a\u8fc7C++\/CUDA\u534f\u540c\u8bbe\u8ba1\u7684\u6781\u81f4kernel\u4f18\u5316\uff0c\u9762\u5bf9512K\u7684KV\u6d77\u6d0b\u65f6\uff0c\u6bcf\u6b21\u68c0\u7d22\u65f6\u95f4<strong>\u4e0d\u5230100\u5fae\u79d2<\/strong>\uff0c\u7aef\u5230\u7aef\u541e\u5410\u91cf\u662f\u539f\u751f\u89e3\u7801\u76843\u500d\u4ee5\u4e0a\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.5 \u5206\u5c42\u8bed\u4e49\u538b\u7f29\uff1aSentenceKV\uff082025\uff09<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bb8\u591a\u9a71\u9010\u7b56\u7565\u5ffd\u7565\u4e86KV\u5411\u91cf\u5185\u90e8\u7684\u6df1\u5c42\u8bed\u4e49\u89c4\u5f8b\u3002\u5f53\u9762\u5bf9\u4e00\u672c\u4e66\u7bc7\u5e45\u7684\u4e0a\u4e0b\u6587\u65f6\uff0c\u5927\u91cfToken\u5176\u5b9e\u5728\u4f20\u9012\u76f8\u4f3c\u7684\u4fe1\u606f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>SentenceKV<\/strong>\u53e6\u8f9f\u8e4a\u5f84\uff0c\u5c06KV\u538b\u7f29\u653e\u5230\u4e86<strong>\u53e5\u5b50\u8bed\u4e49\u76f8\u4f3c\u5ea6\u7684\u5c42\u7ea7<\/strong>\u3002\u5b83\u5728Prefill\u9636\u6bb5\u5c31\u5229\u7528\u7f16\u7801\u5668\u63d0\u53d6\u53e5\u5b50\u7684\u9ad8\u7ef4\u8bed\u4e49\u5411\u91cf\uff0c\u89e3\u7801\u65f6\uff0c\u76f4\u63a5\u5bf9\u6bd4\u751f\u6210\u4e2d\u67e5\u8be2\u7684\u8bed\u4e49\u5411\u91cf\uff0c\u63d0\u524d\u7cbe\u51c6\u9884\u6d4b\u201c\u9700\u8981\u54ea\u4e2a\u53e5\u5b50\u7684KV\u5757\u201d\uff0c\u628a\u6574\u4e2a\u5e9e\u5927\u7684\u4e0a\u4e0b\u6587\u53d8\u6210\u4e00\u53e5\u8bdd\u7ea7\u522b\u7684\u201c\u5185\u5bb9\u5730\u56fe\u201d\u5feb\u901f\u7d22\u5f15\uff0c\u5927\u5e45\u964d\u4f4e\u65e0\u6548\u663e\u5b58\u5197\u4f59\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u52a8\u6001\u6f14\u8fdb\uff1a\u4eceMHA\/MLA\u5230\u5168\u8986\u76d6\u4f18\u5316\u7684\u67b6\u6784\u9769\u65b0<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0e\u6b64\u540c\u65f6\uff0c\u5728\u67b6\u6784\u5c42\u9762\uff0c\u5bf9KV Cache\u7684\u7cbe\u7b80\u4e5f\u5728\u6301\u7eed\u8fdb\u5316\uff0c\u4e0e\u7cfb\u7edf\u7ba1\u7406\u5f62\u6210\u8f6f\u786c\u4e00\u4f53\u7684\u914d\u5408\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>DeepSeek MLA\u518d\u5ba1\u89c6<\/strong>\uff1a\u6211\u4eec\u66fe\u5728Attention\u4e00\u8bb2\u63d0\u5230\uff0cMLA\u57fa\u4e8e\u4f4e\u79e9\u5206\u89e3\u5c06KV\u7684\u663e\u5b58\u538b\u7f29\u6bd4\u63a8\u5230\u6781\u81f4\u3002\u6700\u65b0\u7814\u7a76\u8868\u660e\uff0c\u4e00\u4e2a\u4eceGQA\u8fc1\u79fb\u81f3MLA\u67b6\u6784\u7684LLaMA\u6a21\u578b\uff0c\u5176KV Cache\u76f4\u63a5\u7f29\u51cf\u4e86<strong>92.19%~93%<\/strong>\uff0c\u6a21\u578b\u8d28\u91cf\u4ec5\u5fae\u8dcc\u4e0d\u52301%\uff0c\u63a8\u7406\u901f\u5ea6\u5374\u83b7\u5f97<strong>10.6\u500d\u7684\u52a0\u901f<\/strong>\u3002<\/li>\n\n\n\n<li><strong>MLA\u8fc1\u79fb\u5de5\u5177<\/strong>\uff1a2025\u5e74ACL\u7684\u5de5\u4f5c\u5df2\u5b9e\u73b0<strong>\u96f6\u6210\u672c\u8f6c\u6362<\/strong>\u2014\u2014\u65e0\u9700\u5b8c\u6574\u8bad\u6a21\u578b\uff0c\u4ec5\u7528\u4e00\u5c0f\u90e8\u5206\u6570\u636e\u5fae\u8c03\uff080.6%~1%\uff09\uff0c\u5c31\u53ef\u4ee5\u5c06\u4efb\u4f55\u6210\u719f\u7684Transformer\u6a21\u578b\u65e0\u7f1d\u8fc1\u79fb\u5230\u6781\u4f4eKV\u5360\u7528\u7684MLA\u67b6\u6784\u4e0a\uff0c\u6781\u5927\u964d\u4f4e\u4e86\u8be5\u6280\u672f\u7684\u51c6\u5165\u95e8\u69db\u3002<\/li>\n\n\n\n<li><strong>MoE-nD\uff082026\uff09<\/strong>\uff1a\u8be5\u6846\u67b6\u5728\u6574\u4e2aKV\u7f13\u5b58\u7684\u591a\u7ef4\u5ea6\uff08Token\u9a71\u9010\u3001\u91cf\u5316\u3001\u964d\u7ef4\u3001\u5c42\u5171\u4eab\uff09\u4e0a\uff0c\u901a\u8fc7<strong>\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u6a21\u578b<\/strong>\uff0c\u5bf9\u6bcf\u4e00\u5c42Transformer\u8fdb\u884c\u52a8\u6001\u8def\u7531\uff0c\u5b9e\u73b0\u6700\u201c\u4e2a\u6027\u5316\u201d\u7684KV\u538b\u7f29\u3002\u5728\u51e0\u4e4e\u4e0d\u53d1\u751f\u8d28\u91cf\u635f\u8017\u7684\u524d\u63d0\u4e0b\uff0c\u5c06\u5e9e\u5927\u76841.9GB\u7f13\u5b58\u72c2\u538b\u81f3<strong>136MB<\/strong>\uff08\u538b\u7f29\u7387\u9ad8\u8fbe14\u500d\uff09\u3002<\/li>\n\n\n\n<li><strong>GRACE\uff082026\uff09<\/strong>\uff1a\u5c06KV Cache\u7684\u901a\u9053\u526a\u679d\u5316\u4e3a<strong>\u56fe\u8c31\u4f18\u5316\u95ee\u9898<\/strong>\u3002\u6bcf\u4e2a\u901a\u9053\u4e3a\u201c\u7ed3\u70b9\u201d\uff0c\u901a\u9053\u95f4\u5173\u7cfb\u4e3a\u201c\u8fb9\u6743\u91cd\u201d\uff0c\u901a\u8fc7\u6700\u5c0f\u5316\u6ce8\u610f\u529b\u77e9\u9635\u91cd\u6784\u8bef\u5dee\u6765\u8bc6\u522b\u6700\u4f18\u6700\u5c0f\u5b50\u56fe\uff0c\u5728\u7701\u4e0b60%\u7f13\u5b58\u7a7a\u95f4\u7684\u540c\u65f6\uff0c\u8fd1\u4e4e\u96f6\u6027\u80fd\u9000\u5316\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u603b\u7ed3\u4e0e\u5c55\u671b\uff1a\u6df1\u5ea6\u878d\u5408\u4e0e\u8f6f\u786c\u4e00\u4f53\u7684\u672a\u6765<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 \u6f14\u8fdb\u8def\u7ebf\u603b\u7ed3<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th class=\"has-text-align-left\" data-align=\"left\">\u7ef4\u5ea6<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u65e9\u671f\u65b9\u6848<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u5f53\u524d\u4e3b\u6d41<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u672a\u6765\u65b9\u5411<\/th><\/tr><\/thead><tbody><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u8ba1\u7b97\u6548\u7387<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u91cd\u590d\u8ba1\u7b97O(n\u00b2)<\/td><td class=\"has-text-align-left\" data-align=\"left\">KV Cache (O(n))<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u7a00\u758f\u6ce8\u610f\u529b\u3001CPU-GPU\u878d\u5408<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u663e\u5b58\u7ba1\u7406<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u9759\u6001\u8fde\u7eed\u5206\u914d<\/td><td class=\"has-text-align-left\" data-align=\"left\">PagedAttention\u3001vLLM<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u8bed\u4e49\u5206\u5c42\u3001\u5171\u4eab\u6c60\u5316<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u67b6\u6784\u538b\u7f29<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">MHA\uff08\u500d\u4e58\u5f00\u9500\uff09<\/td><td class=\"has-text-align-left\" data-align=\"left\">GQA\/MQA\u3001MLA (DeepSeek)<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u96f6\u8bad\u7ec3\u8fc1\u79fb\u3001MoE\u8def\u7531\u538b\u7f29<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u5e8f\u5217\u6269\u5c55<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u9884\u7559\u3001\u4e22\u5f03<\/td><td class=\"has-text-align-left\" data-align=\"left\">Lookahead\/TTKV<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u8bed\u4e49\u9884\u6d4b\u3001\u54c8\u5e0c\u6377\u5f84<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u7cbe\u5ea6\u4e0e\u4f4d\u5bbd<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">FP16\/BF16<\/td><td class=\"has-text-align-left\" data-align=\"left\">KV FP8\u538b\u7f29<\/td><td class=\"has-text-align-left\" data-align=\"left\">TurboQuant 3.5-bit \u65e0\u635f<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u591a\u8bf7\u6c42\u590d\u7528<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u7269\u7406\u9694\u79bb\uff0c0\u590d\u7528<\/td><td class=\"has-text-align-left\" data-align=\"left\">Prefix Caching<\/td><td class=\"has-text-align-left\" data-align=\"left\">LMCache\u8de8\u8bf7\u6c42\u5168\u5c40\u590d\u7528<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 \u878d\u5408\u667a\u80fd\u4e0e\u786c\u4ef6\u7684\u672a\u6765<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6211\u4eec\u5df2\u7ecf\u770b\u5230KV Cache\u5728\u7cfb\u7edf\u3001\u7b97\u6cd5\u3001\u67b6\u6784\u4e09\u4e2a\u7ef4\u5ea6\u4e0a\u7684\u5168\u9762\u4f18\u5316\u3002\u800c\u672a\u6765\uff0c\u8fd9\u4e09\u4e2a\u7ef4\u5ea6\u7684\u8fb9\u754c\u6b63\u5728\u52a0\u901f\u6a21\u7cca\u2014\u2014\u8fc8\u5411\u201c\u8f6f\u786c\u4e00\u4f53\u3001\u6df1\u5ea6\u534f\u540c\u201d\u7684\u65f6\u4ee3\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>CPU-GPU\u7684\u6df1\u5ea6\u878d\u5408\u8ba1\u7b97\u6846\u67b6<\/strong>\uff08\u5982 <strong>HybridGen<\/strong>\uff09\u4e0d\u518d\u628aKV Cache\u5f53\u4f5cGPU\u4e13\u6709\u8d22\u4ea7\uff0c\u901a\u8fc7CXL\u7b49\u6280\u672f\u5c06\u5176\u6269\u5c55\u81f3CPU\u3001NVMe\u3001SSD\u7b49\u591a\u4e2a\u5c42\u7ea7\u7684\u5b58\u50a8\u6c60\uff0c\u5b9e\u73b0\u66f4\u7075\u6d3b\u3001\u5927\u5bb9\u91cf\u7684\u6df7\u5408\u63a8\u7406\u3002<\/li>\n\n\n\n<li><strong>\u57fa\u4e8e\u5b58\u5185\u8ba1\u7b97\uff08PIM\uff09\u7684\u76f4\u63a5\u8ba1\u7b97\u67b6\u6784<\/strong>\uff0c\u6709\u53ef\u80fd\u4ece\u7845\u57fa\u5c42\u9762\u5f7b\u5e95\u5378\u6389\u201c\u5185\u5b58\u5899\u201d\u8fd9\u4e00\u7269\u7406\u5929\u82b1\u677f\u3002<\/li>\n\n\n\n<li><strong>\u591a\u4e0a\u4e0b\u6587\uff08SamKV\uff09\u573a\u666f\u4e0b\u7684\u6781\u81f4KV\u590d\u7528\u7b97\u6cd5<\/strong>\uff0c\u5728\u8d85\u5927\u89c4\u6a21\u957f\u6587\u6863\u6c47\u603b\u573a\u666f\uff0c\u672a\u6765\u53ef\u80fd\u4ec5\u4ee5<strong>15%\u7684\u7f13\u5b58<\/strong>\u5f00\u9500\u5b9e\u73b0\u540c\u7b49\u529f\u6548\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u5207\u90fd\u6307\u5411\u540c\u4e00\u4e2a\u7ec8\u6781\u76ee\u6807\uff1a<strong>\u8ba9\u5927\u6a21\u578b\u5728\u771f\u6b63\u7684\u751f\u4ea7\u73af\u5883\u4e2d\uff0c\u4ee5\u6700\u4f4e\u7684\u6210\u672c\u3001\u5904\u7406\u6700\u590d\u6742\u7684\u4efb\u52a1\u3001\u5e76\u8986\u76d6\u6700\u5e7f\u6cdb\u7684\u4eba\u7fa4\u3002<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5199\u5728\u524d\u9762 \u5728\u524d\u9762\u7684\u516d\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u5b8c\u6574\u62c6\u89e3\u4e86Transformer\u7684\u9759\u6001\u67b6\u6784\u2014\u2014\u4eceTokenizer\u5230Embe [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":142,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[44,37],"class_list":["post-141","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mtsb","tag-kv-cache","tag-llm"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/141","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/comments?post=141"}],"version-history":[{"count":1,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/141\/revisions"}],"predecessor-version":[{"id":143,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/141\/revisions\/143"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media\/142"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=141"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=141"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=141"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}